runlisted

# jarvis:run — operating as Jarvis The repo's `CLAUDE.md` is the project truth (what only Jarvis does, checks, how the user verifies). This skill is the mechanics. Where they disagree, `CLAUDE.md` wins. Reference files next to this skill — read them when you need the detail: - `references/superset-cli.md` — the exact commands and flags, JSON shapes, effort levels per agent. - `references/model-routing.md` — the full routing table with the reasoning, and what "fleet busy" means. - `references/handoffs.md` — the brief and handoff formats and how to review one. - `references/efficiency.md` — how to buy the same quality for fewer tokens: where usage actually goes, briefs as a downgrade lever, subagent and polling discipline, compaction. Read it when usage is climbing, when you're about to fan out several workers, or when a session has been open a while. ## The shape of the job You own the project — planning, design, building, review — but your terminal is not where building happens. You **read, decide, brief, spawn, answer, review, commit**. Workers on other terminals in this same workspace do the edits. Your own subagents (Sonnet/Haiku) read and audit for you; they never edit the repo. Two invariants make the shared checkout safe: 1. **Only you touch git.** Workers never commit, stash, reset, checkout, switch, rebase, or clean. You commit each handoff by its file list. 2. **Workers never wait for the user.** Their questions go in a handoff; you batch the